[PATCH] block: fix block_class iteration locking

[PATCH] block: fix block_class iteration locking

[Tejun Heo]

genhd.c used block_class_lock when it wants to iterate over
block_class.devices list.  This is incorrect as block_class_lock has
nothing to do with synchronization around block_class.devices list, so
there's a window, albeit small, where the list can change while it's
being iterated over.  Fix it by using block_class.sem instead.

Signed-off-by: Tejun Heo <tj@kernel.org>
---
 block/genhd.c |   16 ++++++++--------
 1 files changed, 8 insertions(+), 8 deletions(-)

diff --git a/block/genhd.c b/block/genhd.c
index b922d48..b646e7e 100644
--- a/block/genhd.c
+++ b/block/genhd.c
@@ -234,7 +234,7 @@ void __init printk_all_partitions(void)
 	char buf[BDEVNAME_SIZE];
 	int n;
 
-	mutex_lock(&block_class_lock);
+	down(&block_class.sem);
 	/* For each block device... */
 	list_for_each_entry(dev, &block_class.devices, node) {
 		if (dev->type != &disk_type)
@@ -275,7 +275,7 @@ void __init printk_all_partitions(void)
 		}
 	}
 
-	mutex_unlock(&block_class_lock);
+	up(&block_class.sem);
 }
 
 #ifdef CONFIG_PROC_FS
@@ -285,7 +285,7 @@ static void *part_start(struct seq_file *part, loff_t *pos)
 	loff_t k = *pos;
 	struct device *dev;
 
-	mutex_lock(&block_class_lock);
+	down(&block_class.sem);
 	list_for_each_entry(dev, &block_class.devices, node) {
 		if (dev->type != &disk_type)
 			continue;
@@ -311,7 +311,7 @@ static void *part_next(struct seq_file *part, void *v, loff_t *pos)
 
 static void part_stop(struct seq_file *part, void *v)
 {
-	mutex_unlock(&block_class_lock);
+	up(&block_class.sem);
 }
 
 static int show_partition(struct seq_file *part, void *v)
@@ -533,7 +533,7 @@ static void *diskstats_start(struct seq_file *part, loff_t *pos)
 	loff_t k = *pos;
 	struct device *dev;
 
-	mutex_lock(&block_class_lock);
+	down(&block_class.sem);
 	list_for_each_entry(dev, &block_class.devices, node) {
 		if (dev->type != &disk_type)
 			continue;
@@ -560,7 +560,7 @@ static void *diskstats_next(struct seq_file *part, void *v, loff_t *pos)
 
 static void diskstats_stop(struct seq_file *part, void *v)
 {
-	mutex_unlock(&block_class_lock);
+	up(&block_class.sem);
 }
 
 static int diskstats_show(struct seq_file *s, void *v)
@@ -658,7 +658,7 @@ dev_t blk_lookup_devt(const char *name, int part)
 	struct device *dev;
 	dev_t devt = MKDEV(0, 0);
 
-	mutex_lock(&block_class_lock);
+	down(&block_class.sem);
 	list_for_each_entry(dev, &block_class.devices, node) {
 		if (dev->type != &disk_type)
 			continue;
@@ -671,7 +671,7 @@ dev_t blk_lookup_devt(const char *name, int part)
 			break;
 		}
 	}
-	mutex_unlock(&block_class_lock);
+	up(&block_class.sem);
 
 	return devt;
 }
-- 
1.5.4.5

Re: [PATCH] block: fix block_class iteration locking

[Cornelia Huck]

On Mon, 14 Jul 2008 15:06:08 +0900,
Tejun Heo <tj@kernel.org> wrote:

> genhd.c used block_class_lock when it wants to iterate over
> block_class.devices list.  This is incorrect as block_class_lock has
> nothing to do with synchronization around block_class.devices list, so
> there's a window, albeit small, where the list can change while it's
> being iterated over.  Fix it by using block_class.sem instead.
> 
> Signed-off-by: Tejun Heo <tj@kernel.org>

Greg has some patches sitting in his tree that convert those functions
to use class_{for_each,find}_device() instead - and that looks like the
clean solution, especially since the class implementation details will
be moved into a private structure. One worry though...


> @@ -285,7 +285,7 @@ static void *part_start(struct seq_file *part, loff_t *pos)
>  	loff_t k = *pos;
>  	struct device *dev;
> 
> -	mutex_lock(&block_class_lock);
> +	down(&block_class.sem);
>  	list_for_each_entry(dev, &block_class.devices, node) {
>  		if (dev->type != &disk_type)
>  			continue;
> @@ -311,7 +311,7 @@ static void *part_next(struct seq_file *part, void *v, loff_t *pos)
> 
>  static void part_stop(struct seq_file *part, void *v)
>  {
> -	mutex_unlock(&block_class_lock);
> +	up(&block_class.sem);
>  }
> 
>  static int show_partition(struct seq_file *part, void *v)
> @@ -533,7 +533,7 @@ static void *diskstats_start(struct seq_file *part, loff_t *pos)
>  	loff_t k = *pos;
>  	struct device *dev;
> 
> -	mutex_lock(&block_class_lock);
> +	down(&block_class.sem);
>  	list_for_each_entry(dev, &block_class.devices, node) {
>  		if (dev->type != &disk_type)
>  			continue;
> @@ -560,7 +560,7 @@ static void *diskstats_next(struct seq_file *part, void *v, loff_t *pos)
> 
>  static void diskstats_stop(struct seq_file *part, void *v)
>  {
> -	mutex_unlock(&block_class_lock);
> +	up(&block_class.sem);
>  }
> 
>  static int diskstats_show(struct seq_file *s, void *v)

With your patch, you protect against adding/removing devices to/from
the class between _start and _stop. This wasn't done previously, and
won't be done with Greg's patches (which leaves in the
locking/unlocking of block_class_lock). Not sure if I'm missing
something here...

Re: [PATCH] block: fix block_class iteration locking

[Tejun Heo]

Cornelia Huck wrote:
> Greg has some patches sitting in his tree that convert those functions
> to use class_{for_each,find}_device() instead - and that looks like the
> clean solution, especially since the class implementation details will
> be moved into a private structure. One worry though...

Yeah, that sounds nice.

> With your patch, you protect against adding/removing devices to/from
> the class between _start and _stop. This wasn't done previously, and
> won't be done with Greg's patches (which leaves in the
> locking/unlocking of block_class_lock). Not sure if I'm missing
> something here...

That's mostly the point of this patch.  It wasn't protected properly
before.  It would be best if there's a safe iterator (maybe use klist?).

Thanks.

-- 
tejun

Re: [PATCH] block: fix block_class iteration locking

[Cornelia Huck]

On Mon, 14 Jul 2008 20:50:57 +0900,
Tejun Heo <tj@kernel.org> wrote:

> > With your patch, you protect against adding/removing devices to/from
> > the class between _start and _stop. This wasn't done previously, and
> > won't be done with Greg's patches (which leaves in the
> > locking/unlocking of block_class_lock). Not sure if I'm missing
> > something here...
> 
> That's mostly the point of this patch.  It wasn't protected properly
> before.  It would be best if there's a safe iterator (maybe use klist?).

Hm, I don't see how a klist would help here since it still allows to
add/remove devices during walking the file.

What we need is more like a "lock/unlock the device list" operation and
a lockless iterator - that would give the same semantics as your patch.